Skip to content

Onboarding, screen by screen

Guide

This is the whole first run on Hobby, one screen at a time: an account, a workspace, GitHub, one indexed repository, a search, an agent key, and the first finding an agent proposes. The screens come from a new account, Ada Lovelace, connected to a small public repository. Getting started says the same in fewer words.

Open Create an account. Sign up with Google or GitHub, or with a name, an email and a password. Accept the terms, wait for the Cloudflare check to show Success!, then press Sign up.

The sign-up form with a name, an email, a hidden password, the terms accepted and the Cloudflare check passed

A workspace holds your team’s repositories and their one search index. Name it and press Create workspace. Hobby includes one workspace.

Manage workspaces with the name Analytical Engine typed and no workspaces yet

Home is where you search. Until a repository is indexed it says so and sends you to Choose repositories. The plan shows beside your name, bottom left.

Home saying nothing is indexed yet, with a Choose repositories button

Workspace shows your plan, then each setup step and whether it is done, next, or waiting. The next one is Connect GitHub.

The Workspace page with Plan done, Repositories next, Search index waiting and Team done

Open Graph & Index and press Connect GitHub. mux reads only the repositories you grant, and never writes to them. The same page lists the eight tools an agent reaches over MCP.

Graph & Index with GitHub not connected and a Connect GitHub button

GitHub asks which account or organization gets the mux app. Choose the one that owns the code. An account where the app is already installed says Configure instead.

GitHub asking where to install the mux app, listing a personal account and organizations

Choose Only select repositories and pick the repository. The app asks for read access to code and metadata, nothing more. Press Install & Authorize; GitHub sends you back to mux.

GitHub’s install screen with one repository selected and read-only access

Back in mux, tick the repository, keep the default branch or name another, and press Save repositories. Hobby indexes one repository; Pro indexes twenty and Pro+ fifty.

Graph & Index connected to GitHub, with the repository ticked and the main branch

The repository is copied, then its files are cut along their syntax. The row says Copying while it works.

The Index card showing the repository copying and the first index being built

When it is done the row says Current, with the file count and the commit it read. On Hobby, Refresh index brings in new commits when you ask; Pro and Team refresh on every push. If GitHub turns a read away for a moment, mux tries again by itself.

The Index card showing the repository current, with 7 files and the commit

Type a name from the code. On Hobby the box matches names exactly; on Pro and Team you can also ask in your own words. Every passage names the repository, the file and the line, and links to the commit it was read from.

Home with retryDelay searched and exact matches in the source and its test

Open Manage account → Settings → Access. Under API keys, keep principal and role as agent, name the key after the tool that will use it, set an expiry, and press Mint key. Copy the key at once: mux shows it only this one time. It is hidden here.

The Access page with a newly minted agent key, hidden, named claude-code

Give the key to your agent as a Bearer token for https://usemux.com/mcp. For Claude Code, put the variable, not the key, in .mcp.json:

.mcp.json
{
"mcpServers": {
"mux": {
"type": "http",
"url": "https://usemux.com/mcp",
"headers": { "Authorization": "Bearer ${MUX_API_KEY}" }
}
}
}

The MCP guide covers Codex, Claude Code and GitHub Copilot CLI. muxcode signs in from its own window instead; no key is copied by hand.

An agent starts a task with get_context: one packet of evidence tied to the indexed commit, with the files the code graph links to it. When it learns something worth keeping, it proposes a finding with record_context_outcome, naming the files the finding rests on.

A proposed finding waits on the Workspace page for an owner. Nothing reaches another agent until you press Approve. An approved finding comes back with every packet, and moves to each new commit while the files it cites stay the same. Retire takes it out.

The Findings card with one proposed finding and Approve and Retire buttons

Context trace, on the same page, lists each packet: when, which key asked, how many passages, how long it took, which commit it read, and what became of the findings proposed from it. The line on top is the week’s speed.

The Context trace card showing one packet from claude-code and its approved finding